How harnsy works

You already use Claude Code, Codex or OpenCode. Each one works well alone. The trouble starts when you run several. One agent finishes something the next one needs, and you copy it across by hand. You become the post office between your own agents. Your time goes to passing things along, not to deciding.

harnsy links these tools. It runs on your machine. It delivers a message straight into a running session, and the agent reads it in the turn it is already in.

In harnsy you are the boss and the reviewer, not a member of the team. You say what you want. A lead agent splits it into work for the others. They do it and check each other. You step in only when something waits for you: an answer, a permission, or a result to accept or send back.

You talk to the lead, not to every agent. What matters is that the result is what you asked for.

harnsy does not promise work without a person. On hard tasks agents still go off track. What it gives you is a process where your time goes to the lead and to the moments that need you.

A harness is the tool an agent runs in: Claude Code, Codex or OpenCode. harnsy connects the tools you already have. It does not replace them. It uses only the way each tool offers to receive a message. Permissions stay the tool's own.

Each card says the same as the module that uses the word. Come back here when a word is unclear.

A company is the whole picture. It keeps the role templates. A project is a folder you work in. A team belongs to one project folder. A role is a post in the team, with a mandate, goals and a deliverable.

Roles stay. Agents sit on roles. When an agent is replaced, the role stays.

So instructions, versions and written handover notes carry the work, not one session's memory. That is the idea behind «the company stays, the agents change».

Every team has a lead role at the top. The lead hands out work and watches who is idle or stuck. It asks you only what it cannot decide.

Agents hold the roles below. A role can have more than one holder for a while: a successor sits next to the old holder until the handover is done.

The lead is told when:

• an agent is idle with open work;
• an item stalls;
• a holder's context runs high;
• a holder is gone with open work.

The lead does not check on a timer.

Work is tracked as items. An item belongs to a role, not to a session. A successor finds it where the old holder left it.

An item moves New → Taken → In progress → In review → Done. It can also be Blocked while it waits (for you, or for something outside until a date), or Dropped. Nothing moves by itself: someone moves it.

When an agent hands in an item, it writes a summary and evidence. Whoever asked judges the result: accept it, or send it back with a note.

You do not watch the dashboard all day. It calls you: a tray notification, the bell, a wait or a failure. You act and go back to work.

It runs on your machine at 127.0.0.1:7788 and answers only on your machine.

A colour means one thing everywhere. Amber: waits for a person. Green: works. Red: failed, or has no holder. Blue: an action. Every colour also has a word or a shape.

A message is a quick word to one agent, delivered into its running session. An item is work with an owner, a state and a result. A handover is a written note. It lets a new agent take over a role when the old one's context is full.

A message says something. An item gets something done. A handover keeps the work going from one session to the next.

An agent sends a message to another by name. harnsy hands it to the harness the recipient runs in, in that harness's own way. The recipient answers with a reply tied to the first message, so one conversation stays one conversation.

Everything on the bus is recorded. You can see who talked to whom in the Messages tab and the Links graph.

Match each channel to the line that describes it.

If you need a result that someone must judge, make an item. If it is a question or a fact, send a message. If the agent that holds the knowledge is about to leave, write a handover.

Every step an agent takes sends its whole context again, and a context has an end. Each role has a handover threshold. When the holder reaches it, the lead is told.

The work continues because the next agent follows a note. It is not endless memory. harnsy has no memory feature today.

Six steps, from installing harnsy to your first word to the lead. An initiator agent sets the team up with you; it saves nothing without your yes.

harnsy without a key: 1 project, 10 running agents, no other machines. A restore that would go over 10 opens nothing.

A role document is what an agent reads when it starts. It holds the mandate, goals, deliverable, what to read first, who the lead is, the handover threshold, and what the post did before.

A company template gives the base. A project adds its own text on top. A change is saved as a new version, and the holder is told their duties changed.

harnsy does not override a harness's permissions. When an agent asks to run something, or asks you a question, the bell shows it with buttons. You answer there, without opening the agent's terminal. Amber means it is waiting for you.

When an item reaches In review, whoever asked judges it. For your own requests, that is you, on the item's card.

One question. It checks that you can read the dashboard.